Reinforcement learning-based optimal power flow of distribution networks with high permeation of distributed PVs

Owing to the large-scale integration of distributed generation units, traditional distribution networks are gradually turning to active distribution networks, which poses impacts on the distribution networks, including voltage violations, the degradation of power losses and the failures of relay protections. Therefore, higher requirements on the distribution network in terms of planning, control, power quality management, operation and maintenance have been put forward. The traditional OPF algorithms have drawbacks on solving large-scale non-linear problems with dispersiveness and randomness. In order to achieve real-time control of the active distribution network under more complicated network structure conditions, an optimization algorithm based on reinforcement learning is presented and applied to the real-time control processes of active distribution networks. The objective is to minimize power losses, considering some crucial constraints such as network voltages, line load rate and distributed power output. Reinforcement learning, as an artificial intelligence method that does not depend on mathematical models, has great advantages in solving large-scale complex mathematical models. In addition, the control effectiveness and convergences of the method are improved through the expansion of the knowledge matrix, and the adoption of imitative learning, migration learning and risk assessment mechanisms. Based on the test of an improved IEEE 33-bus system under MATLAB R2017a, this paper verifies that the reinforcement learning method has a better effect on the real-time control of the active distribution network in terms of node voltages and network losses. Simultaneously, the method significantly relieves over-limits of voltages, effectively reduces network losses, and has proper real-time characteristics.